Karteikarten
DD303 Language Processing
Karten in diesem Stapel (20)
- 1What are the main differences between spoken & written word recognition?Spoken:
>Some aspects innate
>Primary
>Need to segment words
>Extended in time but each word exists only briefly
Written:
>Learned (but likely we use mechanisms for spoken lang recognition)
>Secondary
>Segmentation is done by white space on page
>Extended in space but words are permanent
Suggests, different models needed to explain each - 2Segmenting spoken words - 2 methods?Pre-lexical (bottom-up):
>Characteristics indicate boundary
>English - stressed 1st syllable - Cutler & Carter
>Different for other languages
>Cutler & Norris - word spot task, embedded in nonsense syllables
>BUT must be alternative for words that break rules (eg guitar, yacht)
Lexical models (top-down):
>Phonological representation (knowledge of what word sounds like)
>Recognise each word & predict boundary
>Match sounds against store (mental lexicon)
>Marslen-Wilson & Welsh - problems w/ short words? - 3Saffran et al (1996)?>Babies - head turning experiment
>Artificial language
>Id words in 2 mins
>Suggests implicit learning ability via co-occurrence of syllables & statistical properties - 4Parallel activation in spoken word recognition?Simultaneous hearing & evaluation
Content processed before end of word
Marslen-Wilson et al ('78 & '87) - Cohort model
Word start - initial cohort
Word continues - list of matches reduces
Single word recognised - uniqueness point
Cross modal priming & semantic similarity
Gaskell & Marslen-Wilson ('02) - # words simultaneously activated limited; suggests priming effect weakens - 5Lexical competition - connectionist modelseg TRACE - McCelland & Elman
3 levels - 1) Phonetic features;
2) Phonemes; 3) Word
Activation - bottom-up - nodes representing phonemes activated when hear a word
When all phonemes for a word are activated, activation spreads to the word
Amount depends on closeness of match
Phonemes c, o, en, f = "confess" & "confetti"
Inhibition @ word level = competitive part of model
Support - McQueen et al ('84) - segmentation via word spotting - 6Visual word recognition - IAC modelMcClelland & Rumelhart ('81)
Top-down
3 levels: 1) visual features; 2) letters;
3) words
Explains word superiority effect (letters recognised faster if in string ie word)
BUT effect can be explained with no top-down feedback using spelling representations in mental lexicon (Grainger & Jacobs '94)
Top-down feedback needed for word recognition? = Contentious - 7Interaction of spoken & written words - DRC modelColtheart et al ('01)
Dual Route Cascade Model
Assembled (rule-based, regular words) vs addressed (lexical, irregular words) phonology
Speed of written word naming = assess which route used
Regular faster than irregular (only low frequency)
Regular use either route, irregular only one route - gives advantage for regular words
Phonology important even in silent reading (e.g. rose, rows) (Van Orden, 1987)
Glushko ('79) neighbouring word properties impact recognition rate - eg save, wave, gave - have
BUT : Jared ('02) - word naming times affected much more by consistency than regularity - 8Eye movements in readingActive - coordinated
Saccades
Fixations - 200ms (dependent on word frequency - Rayner & Duffy)
O'Regan & Jacobs - recognise words faster @ optimal viewing position - mid-word
Shillock et al - info content greatest mid-word; less info at word ending
~10% saccades backwards - more with low predictability/ambiguity
Text is present 'ahead' as well as 'before' - parallel process?
Flanker effect = yes; normal reading evidence = no - 9What is the mental lexicon?Store
Semantic r'ships - how meanings related
Semantic content - meaning
Phonological
Orthographic - 10Morphology?Size of units in lexicon
Morpheme - smallest meaningful sub-unit
Inflections - plural; tense (-ed, -ing)
Derivations - change word type (weak>weakly=adjective>verb)
Irregular forms - mouse>mice
Share morphemes - depart, department = unrelated
2 cognitive approaches:
Full-listing: all words stored in entirety
Decompositional: Taft & Forster - words broken into morphemes & stored w/ links
Support:Marslen-Wilson et al ('94) - words w/ same morpheme prime strongly; faster processing prime & target
Related meanings (cruel>cruelty) NOT shared prefix (casual>casualty) - 11Accessing Meaning - Semantic Representations?Type of info accessed when word recognised?
Spreading activation models (eg Collins & Loftus) - Connectionist: links = semantic r'ship; meanings related
Vs
Featural theories - meanings=set of concepts stored in Lexicon
(Both models poorly specified & data can often be interpreted to fit either) - 12Semantic priming?Semantic priming experiments: e.g. bread>butter, cheddar>cheese
Associative links between prime & target
Lucas ('00) - Non associative (e.g. horse and sheep, much weaker link & data robust)
BUT evidence different visual & perceptual properties resulting in some degree of priming (e.g. size/shape) - 13Semantic ambiguity?Homonyms (e.g. bank – river vs money)
Context allows selection of appropriate meaning, but how?
>Autonomous (all meanings accessed then select contextually appropriate one)
>Interactive (context rules out some meanings before they are fully activated)
>Swinney ('79) supported autonomous model - Regardless of bias, both possible words primed; BUT if delay (1 sec) between stimuli, contextually appropriate word primed
>Lucas ('99) supported interactive model - more priming re appropriate meanings then inappropriate - 14Processing sentences?Sentences are novel
Implies process has to be constructive, rather than just recognition
Parsing = deciding each word’s syntactic (grammatical) role in the sentence to understand the sentence - 15Incremental model of parsing?Syntactic structure built incrementally as each word encountered
Most evidence for this one
Tyler & Marslen-Wilson ('77)
Ambiguous phrase "landing planes"
Preceding context determines if landing = adjective or verb
Speed of response dependent on prior context
Supports incremental parsing
Incompatible with delayed model - it wouldn't predict an effect of appropriateness, as waiting to the end of the sentence, to parse - 16Garden path model of parsing?"The horse raced past the barn fell"
Fraizer ('79)
Influential
>Incremental & autonomous
>Misinterpretations constantly corrected
>Listener might be ‘led down the garden path’
>Serial model – parser makes only one potential parse at a time
>Parsing based on info from sentence itself, not context (syntactic info only) - 17Constraint-based models of parsing?MacDonald et al ('94)
Parsing = parallel & interactive
Multiple interpretations evaluated in parallel
No autonomous component
Info from lexicon added during evaluation process
Evidence:
Multiple meanings of ambiguous words briefly activated
Frequency of meanings (in the language) determines activations of alternatives
Biasing contexts increases activation of an alternative - 18Is parsing autonomous?For:
Ferreira & Clifton ('86) - eye-tracking re verb in "Defendant examined..." - garden path effect found
Against:
Trueswell et al - garden path sentences could be affected by meaning: same contexts as F&C not as constraining as their example
Parsing can be affected by semantic plausibility of various parses of the system - 19Constraints on parsing?Ambiguity reduced by intonation - listeners use this (Warren)
Tannenhaus et al ('95)
"Put the apple on the towel in the box”
environmental info used to resolve ambiguity & reduce/remove garden path effect - 20Starr & Rayner ('01) - key points?>Eye movements in reading claimed to reflect high level cognitive processes
>Studies showing effects of parafoveal (right of fixation) info uptake support models of parallel processing in reading.
>Challenges serial spotlight models of attention in reading
>Frequency & predictability effects on fixation times suggest top-down cognitive influences of lexical & comprehension processes
>Supports models of integrated & parallel processing (MacDonald et al '94)
Informationen
- Beschreibung
- DD303 Ch8 Language Processing
- Erstellt von
- Ken Adams
- Aufrufe
- 152
- Tags
- Bewertung